The Concept
Every model writes
the same way, every time.
GPT-4 has sentence length distributions, punctuation preferences, hedging patterns, transitional phrase tendencies. Claude uses specific discourse markers. Gemini structures paragraphs in identifiable ways. These patterns persist across topics, languages, and styles — because they are baked into the model weights themselves.
Humans can't see these patterns. Statistical models can. The Uncanny Valley of Text is a forensic tool that reads the fingerprint invisible to the naked eye and returns a verdict: which model, which version, which prompt archetype produced this text.
Statistical fingerprint analysis · Sample text
GPT-4o · 94% match
Claude 3 Sonnet · 61% match
Gemini 1.5 Pro · 28% match
Llama 3 70B · 41% match
Verdict → GPT-4o · High confidence (94%) · Prompt archetype: Instructional / Expository · System prompt detected: likely includes "be thorough"
Features
What it detects
Model Attribution
Identifies which AI model produced a piece of text — GPT-4, Claude, Gemini, Llama — with confidence scores across each candidate.
Version Detection
Models change with version updates. The fingerprint evolves — and this tracks it. GPT-4 vs GPT-4o have measurably different stylistic distributions.
Prompt Archetype Inference
Infers the likely prompt structure that produced the output — instructional, conversational, creative, system-prompted — from the output's statistical shape.
Feature Visualiser
Shows the exact stylometric features driving the verdict — sentence length, hedging, discourse markers — making the decision fully interpretable.
Continuous Model Updates
As new models release, the fingerprint library expands. The system is built to be extended — not a fixed classifier, a living forensic database.
Research Mode
Exposes raw feature vectors and statistical distances for researchers studying model behaviour, AI safety, and content authenticity.
Stack
How it's built
Python
Core Analysis
NLTK / SpaCy
Stylometry
HuggingFace
Embeddings
Scikit-learn
Classification
Plotly
Visualisation
FastAPI
Analysis API
React
UI
PostgreSQL
Fingerprint DB
Currently in development
Feature extraction pipeline built · Training data collection in progress · Initial classifier at 78% accuracy on 4 models